search

open-mmlab / mmpretrain

OpenMMLab Pre-training Toolbox and Benchmark
https://mmpretrain.readthedocs.io/en/latest/
Apache License 2.0
3.38k stars 1.05k forks source link

How do you train SEResNet? #302

Closed zhongqiu1245 closed 3 years ago

zhongqiu1245 commented 3 years ago

Hi, dear authors! Sorry to bother you. When you were training SE-ResNet50, did you start training from scratch or did you use pre-training weights loaded with resnet50(load_from = resnet50 or pretrain = resnet50)? Thank you!

mzr1996 commented 3 years ago

Hello, we have uploaded our original train config, you can simply check or use it.

zhongqiu1245 commented 3 years ago

Hi @mzr1996 The link you uploaded is invalid(404). Can you upload it again? Thank you!

zhongqiu1245 commented 3 years ago

I have seen your config.py about SEResNet before, and it looks like you are training from scratch, because load_from=None in the config file. Am I right?

mzr1996 commented 3 years ago

I have seen your config.py about SEResNet before, and it looks like you are training from scratch, because load_from=None in the config file. Am I right?

I have updated the link. Right, it's trained from scratch if the load_from and pretrained config is None

zhongqiu1245 commented 3 years ago

Thank you for your reply！ By the way, what does the value of step in lr_config mean? How to calculate these values?

zhongqiu1245 commented 3 years ago

Any news？

mzr1996 commented 3 years ago

Any news？

Hello, the step key means at which epoch, we will decay the learning rate, and the default decay ratio gamma=0.1. For example, in config

lr_config = dict(policy='step', step=[50, 100])

After the 50th epoch, the learning rate will become 0.1 × Lr and after the 100th epoch, the learning rate will become 0.01 × Lr